Artificial intelligence promises to reshape industries and economies. Yet its physical demands now rival those of entire nations. A report from the United Nations University Institute for Water, Environment and Health lays out the scale in stark terms. Global data centers consumed 448 terawatt-hours of electricity in 2025. That figure matches the residential needs of 1.3 billion people in Sub-Saharan Africa for 2.6 years. AI accounted for roughly one-fifth of that total.
By 2030 those centers could draw 945 terawatt-hours. The jump would push them to nearly 3 percent of worldwide electricity use. Carbon emissions tied to that power would hit 399 million tonnes of CO₂ equivalent. Offsetting such output would require growing 6.7 billion tree seedlings for a decade. That equals twice the total number of trees in the United Kingdom. United Nations University Institute for Water, Environment and Health.
Water use tells an even more urgent story. The same 2025 electricity demand carried a water footprint of 4.5 trillion liters. Enough to fill 1.8 million Olympic-sized swimming pools. Or to meet the basic annual domestic needs of more than 600 million people in Sub-Saharan Africa. Projections for 2030 show AI-related water consumption alone reaching 9.3 trillion liters. That volume would satisfy the yearly household requirements of 1.3 billion people across the region. The land footprint follows suit. More than 14,500 square kilometers by decade’s end. Twice the size of the Jakarta metropolitan area and its 32 million residents.
These numbers come at a moment when tech giants report rising emissions despite long-standing climate pledges. Google saw its total carbon emissions climb 25 percent year-over-year in its latest sustainability update. Amazon recorded a 16 percent increase. Both companies pointed to the surge in AI data-center construction and the electricity required to run advanced models. “The environmental footprint of the data centers that power AI is growing, creating a dual challenge,” Google’s report stated. The Guardian.
But the problem runs deeper than electricity bills and emission tallies. Cooling those dense racks of GPUs demands vast quantities of water. A typical 100-megawatt data center in the United States can consume about 2 million liters per day for cooling alone. That equals the daily water use of roughly 6,500 households. Indirect consumption tied to power generation often dwarfs the direct figure. The International Energy Agency estimates global data-center water use at 560 billion liters annually now, with potential to reach 1.2 trillion liters by 2030 as chips run hotter. Bloomberg.
In drought-prone regions the tension sharpens. Communities near planned facilities in Mexico, Uruguay and parts of the American Southwest have watched local water tables drop while data centers secure priority access. Phoenix-area centers already pull hundreds of millions of gallons daily. One analysis found nearly 60 facilities there drawing around 177 million gallons a day. Plans for expansion could multiply that demand. Local officials in Georgia have fielded proposals for new centers that would consume millions of gallons more than entire counties currently use. The pattern repeats. Benefits accrue to distant users. Burdens land on nearby residents.
Professor Kaveh Madani, director of the UNU-INWEH and 2026 Stockholm Water Prize laureate, frames the issue clearly. “This report is not a case against artificial intelligence. It is a call for using it responsibly and addressing its unintended impacts proactively to make it sustainable and equitable.” He and his co-authors, including lead researcher Dr. Miriam Aczel, emphasize that focusing solely on carbon misses the trade-offs. Switching from coal to bioenergy might cut emissions 70 percent yet increase water use thirtyfold and land requirements a hundredfold. “What surprised us most is how often the choices that look greenest from a carbon perspective end up worse for water or for land,” Aczel said. UNRIC.
Inference, not training, now dominates the energy ledger. Once models deploy, the repeated processing of user prompts accounts for 80 to 90 percent of total consumption. ChatGPT alone handles some 2.5 billion prompts each day. That activity alone equates to roughly 383 gigawatt-hours per year. A single AI-generated image can consume as much electricity as a 10-watt LED bulb running for 17 minutes and roughly 29 milliliters of water. Generating a short AI video multiplies the impact dramatically. Up to 4.1 liters of water per clip. The equivalent of two days’ basic need for one person. These per-query costs remain invisible to most users. Defaults for image resolution or output length drive much of the demand.
Recent studies reinforce the trend. One analysis published late last year found AI systems in 2025 produced carbon emissions comparable to all of New York City. Between 33 million and 80 million tonnes of CO₂. AI-related water consumption exceeded previous estimates for all data-center activity and surpassed global bottled-water demand. Alex de Vries-Gao, the researcher behind the work, called it the first comprehensive attempt to quantify AI’s water impact. The Guardian.
Even optimistic efficiency gains may not deliver net relief. The Jevons paradox appears at work. Cheaper, faster AI encourages far more usage. Improved chips lower the energy per task yet overall volume rises. Dr. Mir Matin, another UNU-INWEH co-author, notes the geographic mismatch. “If you map where data centres are built against where water stress is worst, you tend to see the same regions.” The United States and China host 90 percent of specialized AI compute. Only 32 countries have any meaningful share. More than 150 lack sovereign capacity entirely. This concentration widens both digital and environmental divides.
Hardware adds another layer. By 2030 AI infrastructure could generate 2.5 million tonnes of electronic waste annually. The equivalent of 250 Eiffel Towers. Extraction of rare minerals for chips and servers often occurs in jurisdictions with limited oversight, shifting environmental and social costs to the Global South. A February 2026 analysis found that 74 percent of industry claims about AI delivering climate benefits remain unproven. Researchers could not identify a single verified case of consumer generative tools producing substantial, measurable emissions reductions. Greenpeace.
Local resistance has grown. In Ireland data centers consumed 21 percent of national electricity by 2023, prompting a pause on new approvals in Dublin until 2028. Similar debates play out in Australia, South Africa and across the American West. Officials in the United Kingdom recently revised upward their estimates of AI data-center carbon impact by more than a hundredfold. The revised range for the next decade runs from 34 million to 123 million tonnes of CO₂. Up to 3.4 percent of the country’s projected total emissions. The Guardian.
A study released just days ago added a further complication. AI-driven productivity improvements in oil, gas and coal production appear likely to enable more emissions than applications in renewable energy can offset. Across dozens of scenarios the net effect increased annual carbon pollution by 0.47 to 1.8 gigatonnes. That represents 1 to 5 percent of the global energy sector’s yearly output. For emissions to break even, renewable productivity gains would need to outpace those in fossil fuels by at least four times. The Guardian.
The UNU-INWEH team offers no blanket condemnation. Instead it proposes six principles to guide development: transparency in measurement, efficiency by design, equity and environmental justice, full lifecycle responsibility, global cooperation and sustainable use. Governments should integrate AI infrastructure into national energy, water and land-use planning. Companies must report combined carbon, water and land footprints rather than carbon alone. Operators need to weigh siting decisions against local resource stress. Users and developers alike should favor fit-for-purpose models over default maximum settings. Investors can treat these footprints as material risks.
Professor Tshilidzi Marwala, rector of the United Nations University, ties the issue to broader governance. “The concentrated development of AI infrastructure in the privileged areas of the world is creating a large digital divide that poses profound challenges in the equitable development of AI. Whether it does so equitably is now a governance question.”
Tech executives have begun to acknowledge the pressure. Some firms explore advanced cooling techniques, renewable power purchase agreements and even siting new facilities near abundant hydroelectric or geothermal resources. Yet the pace of model scaling continues to outrun efficiency improvements. A Lawrence Berkeley National Laboratory analysis projects U.S. data-center electricity demand could double or triple by 2028. Direct water consumption may quadruple in the same period.
So the tension persists. AI already influences drug discovery, climate modeling and grid optimization. Those applications carry genuine potential to reduce emissions elsewhere. But without deliberate constraints on inference volume, transparent accounting across all three footprints and equitable distribution of both costs and benefits, the technology risks amplifying the very problems it seeks to solve. The data now speak with clarity. Water tables drop. Grids strain. Emissions climb. The window for embedding responsibility into AI’s expansion narrows with each additional terawatt-hour and trillion liters consumed.